Faadu

Data analysis - processing and imaging
Minimum-phase signal calculation using the real cepstrum

Adrian D. Smith, Robert J. Ferguson
ABSTRACT
The concept of minimum phase in geophysics is an important one, especially for pro-
cesses such as statistical deconvolution which assume the condition in the source wavelet.
We wish to have an alternative method to the Hilbert transform to convert a signal of arbi-
trary phase to its minimum phase equivalent, while retaining the same amplitude spectrum.
We implement a minimum-phase reconstruction based on the real cepstrum developed for
a finite-impulse response (FIR) filters by treating the signal as a filter. We demonstrate that
the algorithm is able to handle signals with ill-conditioned amplitude spectra and still give
minimum-phase outputs through analysis of pole-zero plots, along with a simple decon-
volution test. We also introduce two metrics: the Pole-Zero Ratio (PZR) and Pole-Zero
Distance (PZD) as potential quantitative descriptions of how close a signal is to being min-
imum phase.
INTRODUCTION
The concept of minimum phase signals plays a very important role in seismic data
processing. One particular class of processing algorithms requiring the seismic source sig-
nature under study be minimum phase are statistical or "blind" deconvolution methods,
both stationary (Robinson, 1967; Robinson and Treitel, 1967) and non-stationary (Mar-
grave et al., 2011). These processes are based on the convolutional model of the Earth,
whereby a source wavelet is convolved with a reflectivity function to generate a recorded
seismic trace (Yilmaz, 2001). Deconvolution aims to remove the signature of the source
from the trace, leaving reflectivity. This reflectivity represents subsurface structure - the
imaging of which is the goal of seismic reflection surveys.
The definition of "minimum phase" or "minimum phase delay" comes originally from
filtering theory in digital signal processing (Lamoureux et al., 2011). The classical mathe-
matical definition of a minimum-phase system is one where all of the poles and zeroes of
its rational transfer function in the Z-transform domain lie inside the unit circle on the com-
plex plane (Karl, 1989; Oppenheim and Schafer, 2009). There is a discrepancy between
this definition and the definition used in geophysics, which is that minimum-phase fil-
ters/systems have poles and zeroes outside the unit disk (Lamoureux and Margrave, 2007a;
Lamoureux et al., 2011). This is due to the difference in the definition of the powers of z
in the Z-transform (Margrave, 2013). We use the classic definition in this report.
The definition of a minimum-phase signal is physical in nature, defined as a signal

where the energy in the signal is concentrated near the front of the signal (Lamoureux et al.,
2011). This definition is useful for seismic processsing, as experimentation has shown
that impulsive seismic sources, dynamite for example, have this property (Sherwood and
Trorey, 1965; Ziolkowski and Bokhorst, 1993). We cannot use the classical system def-
inition of poles and zeroes to precisely describe minimum-phase signals, as one cannot
assume the spectra of general signals are always well-represented by rational functions
(Lamoureux and Margrave, 2007a). Also, real signals are bandlimited, with an assumed
CREWES Research Report — Volume 26 (2014) 1

Smith & Ferguson
amplitude spectrum outside that range to be zero. This causes computational issues when
calculating a minimum-phase equivalent using pre-existing methods (Lamoureux and Mar-
grave, 2007a). We propose however to think of signals as digital filters, classify them as
mixed or minimum phase using the classic definition of poles and zeroes. While inaccu-
racies may be introduced by doing so, we feel that using the poles and zeroes of a signal
is useful, especially in cases where we want to describe how "close" a signal is to being
minimum phase and how that may impact processing results, such as deconvolution.
Motivation
Although we do not have direct control over seismic sources to ensure that they gen-
erate a mathematically perfect minimum-phase source, the assumption seems to hold well
enough with seismic data in order for deconvolution to work (Ziolkowski and Bokhorst,
1993). However there are other datasets based on measuring reflected electromagnetic ra-
diation that are similar to seismic in nature, both geophysical such as in georadar (GPR)
(Jol, 2008; Gurbuz et al., 2009), and medical imaging (Fear et al., 2002; Bourqui et al.,
2012), where we in theory have better control over the source. Should we want to use
seismic-based processing methods requiring a minimum-phase source, we need to be able
to precondition a source pulse in order to generate the desired source wavelet. At some
point, we would require the ability to convert an arbitrary mixed-phase wavelet to a min-
imum phase equivalent, with the same amplitude spectrum. This is because such radar
signals tend to be defined only in terms of their desired amplitude spectrum characteristics,
with little concern give to phase.
Lamoureux and Margrave (2007c) and Lamoureux and Margrave (2007b) describe
methods and give examples of computing more generally the minimum phase version of
any band-limited signal using a stable criteria, using truncation of infinite-impulse response
(IIR) filters. In this report, we aim to accomplish two things: 1) implement and test an al-
ternative algorithm to convert a signal to its minimum phase equivalent and 2) introduce
simple metrics for quantifying the "minimum phaseness" of a signal.
The algorithm we implement was developed by Pei and Lin (2006) as a method to de-
sign an arbitrary length minimum-phase finite-impulse response (FIR) filter from a mixed-
phase input. It allows us to calculate the minimum-phase equivalent sequence of an arbi-
trary mixed-phase input with the output having the same amplitude spectrum of the input.
This algorithm has several advantages over other ones:
• The difficulty of root finding (especially in high-order filters) such as in the method
of Herrmann and Schuessler (1970) is avoided. This allows us to calculate the
minimum-phase equivalent of longer input signals than other methods.
• The cumbersome process of phase unwrapping as required in the method of Mian
and Nainer (1982) is avoided.
• A similar method developed Stathaki and Fotinopoulos (2001) requires that the input
be linear-phase, while the input for the method under investigation (Pei and Lin,
2006) can be of general mixed-phase.
2 CREWES Research Report — Volume 26 (2014)

The algorithm extends from the previous work of Mian and Nainer (1982) and avoids
the phase unwrapping problem by using the real cepstrum (Pei and Lin, 2006). The method
requires only two fast Fourier transforms (FFT)’s and a recursive procedure to find the
minimum-phase equivalent of a signal from its real cepstrum. Although this algorithm is
designed for sequences/filters, signals under study are represented as discrete signals, and
as such we apply the reconstruction algorithm to them.
We now cover several important points concerning the cepstrum and how they relate
to the algorithm designed by Pei and Lin (2006), which we give the name minrceps. We
then describe the algorithm itself and provide examples of the performance of the minrceps
algorithm using both pole-zero plots and a simple deconvolution test. To conclude, we
attempt to quantify the degree of a signal’s "minimum phaseness", and what that implies for
the accuracy of geophysical processes relying on the minimum-phase source assumption,
such as deconvolution.
THE CEPSTRUM
The power cepstrum of a signal was defined in Bogert et al. (1963) as the squared mag-
nitude of the inverse Fourier transform of the natural logarithm of the squared magnitude
of the Fourier transform of a signal. This is given as follows:
P C = |F −1 {ln |F{x(n)}|2 )}|2 , (1)
where P C denotes the power cepstrum, x(n) is a real sequence, ln the natural loga-
rithm, and F and F −1 denote forward and inverse Fourier transforms respectively. Let
X(eiw ) represent the Fourier transform of x(n).
From this, the sequence’s complex cepstrum ĉ(n) and its real cepstrum ĥ(n) can be
defined as
Ĉ(eiw ) = ln X(eiw )

h i
iw i arg[X(eiw )]

= ln X(e ) e

= ln X(eiw ) + i arg[X(eiw )] (2)

Ĥ(eiw ) = Re{Ĉ(eiw )} = ln X(eiw ) (3)
ĉ(n) = F −1 {Ĉ(eiw )} (4)
ĥ(n) = F −1 {Ĥ(eiw )} (5)

Smith & Ferguson
(Oppenheim, 1965). Note in Equations 2 and 4 when calculating the complex cepstrum,
we need to compute the natural logarithm of a complex number. The imaginary part of the
complex logarithm must be continuous and without linear-phase term to avoid ambiguity
(Pei and Lin, 2006).
Minimum/maximum phase sequences and their complex cepstra
There are two useful properties from Oppenheim and Schafer (2009) that can be used
to describe the special relationship between minimum and maximum phase sequences and
their complex cepstra:
1. If x(n) is a minimum-phase sequence, its complex cepstrum ĉ(n) will be a causal

sequence - that is, ĉ(n) = 0 for n < 0.
2. If x(n) is a maximum-phase sequence, its complex cepstrum ĉ(n) will be an anti-
causal sequence - that is, ĉ(n) = 0 for n > 0.
Relationship between minimum-phase sequence and its complex cepstrum
An arbitrary sequence x(n) and its complex cepstrum ĉ(n) are related by an implicit
recursive relation from Oppenheim and Schafer (2009) as

ĉ(0)
e ∞ ,
 if n = 0
x(n) = X k (6)
ĉ(k)x(n − k) if n 6= 0
n


k=−∞
From this, if x(n) is a finite minimum-phase sequence, the summation in Equation 6 is

reduced to finite terms as

ĉ(0)
e
 if n = 0
n
x(n) = X k (7)
 ĉ(k)x(n − k) if n > 0

k=0
n
Reconstruction of a causal sequence from its even part
If x(n) is a causal sequence and xe (n) = [x(n) + x(−n)]/2 is the even part of x(n),
then x(n) can be recovered from xe (n) as

0 if n < 0

x(n) = xe (n)u+ (n), where u+ (n) = 1 if n = 0 (8)

2 if n > 0

Fourier transform pair between time and frequency domain
Let xe (n) and x0 (n) be the even and odd parts of a sequence x(n) respectively. If we
take the real and imaginary parts of the Fourier transform of x(n), we have the following
relations:
xe (n) = F −1 {Re X(eiw ) }

(9)
xo (n) = F −1 {iIm X(eiw ) }

(10)
Relationship between complex cepstrum and real cepstrum
If x(n) is a real-valued sequence, its corresponding complex cepstrum is also real-

valued according to the recursive relation given in Equation 6. As this assumption holds
true for the most part, we denote ĉe (n) as the even part of ĉ(n) - which is the complex
cepstrum of x(n). If we substitute X(eiw ) with Ĉ(eiw ) = ln [X(eiw )] into Equation 9 we
get:
ĉe (n) = F −1 {Re[Ĉ(eiw )]} = F −1 {ln X(eiw )}

(11)
Continuing, we use the definition of the real cepstrum in (3) and (5), we get
ĉe (n) = F −1 {Ĥ(eiw )} = ĥ(n) (12)
This result shows that the real cepstrum ĥ(n) is actually the even part of the complex
cepstrum ĉ(n) (Benesty et al., 2008).
The properties discussed throughout this section are required for the development of
the minrceps algorithm. We now describe implementation of the algorithm.
IMPLEMENTATION OF MINRCEPS ALGORITHM
The properties of the cepstrum mentioned in the previous section section to construct a
minimum-phase sequence. Given a mixed-phase real valued input signal x(n), we want to
generate a signal xmin (n) which is minimum phase and has the same amplitude spectrum
as the input

Xmin (eiw ) = X(eiw ) . (13)

Smith & Ferguson
Since the real cepstrum is uniquely determined by the magnitude component of a se-
quence’s frequency response (Relations 3, 5 and 13), two sequences with the same am-
plitude spectra must have the same real cepstra (Benesty et al., 2008; Pei and Lin, 2006).
Therefore,
ĥ(n) = ĥmin (n). (14)
Recall from earlier, ĉmin (n) is a causal sequence and ĥmin (n) is the even part of ĉmin (n)
(Benesty et al., 2008). We can use this relationship and equation 8 to reconstruct ĉmin (n)
from ĥmin (n). After we have found ĉmin (n), we can get xmin (n) from equation 7 (Op-
penheim and Schafer, 2009). xmin (n) represents the minimum-phase sequence that we are
interested in calculating.
There is a need to avoid zeroes on the unit circle. This problem can be overcome by
first multiplying the input sequence x(n) with the following exponential sequence:
(
αn x(n) n = 0, . . . , N − 1
xα (n) = with L 8N, (15)
0, n = N, . . . , L − 1
where α < 1 and α ∼ = 1. This step causes the radius of its zeroes scaled down by the
factor α, moving them slightly inside the unit circle (Pei and Lin, 2006). Also, even though
the sequence is finite, its cepstrum sequence is still infinite (Oppenheim and Schafer, 2009).
To reduce the aliasing effect, we need to add trailing zeroes to αt x(n), as in Equation 15
Minrceps algorithm
1. Choose a value α ≤ 1 and α ∼

= 1 to move the zeroes slightly inside the unit circle as
in Equation 15.
2. Perform an L-point FFT on xα (n) = αn x(n), where n = 0, 1, . . . , (N − 1), to get
Xα (k), k = 0, 1, . . . , (L − 1), where L 8N .
3. Perform an IFFT on ln |Xα (k)| to get ĥα (n), which is equal to ĥα,min (n).
4. Construct ĉα,min (n) from ĥα,min (n) using Equation 8.
5. Calculate xα,min (n) from ĉα,min (n) using the recursion formula in Equation 7.
6. Rescale xα,min (n) to get xmin (n) = xα,min (n)α−n .
Pei and Lin (2006).

TEST EXAMPLES
We implemented the minrceps algorithm presented in the previous section using Oc-
tave, a MATLAB-like environment (Eaton et al., 2013). We test four cases to demonstrate
the suitability of the algorithm. We plot a comparison of the input and output signals, as
well as their amplitude and phase spectra respectively. We use pole-zero plots to illustrate
weather or not the signal in question is minimum-phase using the classic system definition.
In this regard, we consider the signal values to represent filter coefficients of a system,
comprising the numerator of the rational transfer function. We set the denominator to one,
which results in a single pole at the origin. Finally, we compute the absolute difference
between the minimum phase reconstruction using our "minrceps" method and a recon-
struction of the output using the Hilbert transform of the amplitude spectrum (Quatieri and
Oppenheim, 1981).
Our four test cases are summarized as follows:
Case A: Minimum phase wavelet

Case B: Mixed-phase wavelet with linear phase
Case C: Mixed-phase wavelet with linear phase and a number of isolated zeroes in the
amplitude spectrum
Case D: Mixed-phase wavelet with linear phase and an interval of zeroes in the amplitude
spectrum
For the Case A, we generate a minimum-phase seismic wavelet using the wavemin func-
tion in the CREWES toolbox (Margrave, 2001). The input parameters are given in Table 1.
Cases B-D have the same parameters as Case A, and we use the amplitude spectrum from
Case A as a starting point for all four cases. In Case B, we generate a linear phase spectrum
and combine it with the amplitude spectrum from Case A to generate our input signal. In
Case C, we start with the linear phase spectrum from Case B and we take the amplitude
spectrum from Case A, zeroing out five values in non-adjacent locations to create isolated
zeros. Case D is similar to Case C, but this time we zero out an interval of five values in
the amplitude spectrum to generate our input signal. For the minrceps algorithm, we set
the values of the parameters α and L to be 0.9999 and 24 respectively.
Parameter fdom dt ns lt
Value 30Hz 4ms 128 samples 0.508s
Table 1. Parameters of input wavelets used to test minrceps algorithm
Minimum-phase input (Case A)
This first test case represents the "do nothing" case - we would expect that the minrceps
algorithm would not modify a minimum-phase input. In Figure 1 we see that the output

Smith & Ferguson
signal, amplitude spectrum, and phase spectrum are nearly identical. Looking at pole-zero
plots of the input and output in Figure 2, we see little to no change in the location of zeroes.
Finally, looking at the difference between the minrceps output and a Hilbert transform
reconstruction in Figure 3, we see that there is a very small absolute difference between
the two, with a maximum percentage difference on the order of magnitude of 1%. This is
likely due to the bandlimited nature of the input, and possibly accumulation of rounding
errors as well. We conclude that the algorithm sucessfully passes the "do nothing" case.
(a)
Input Signal
Minimum Phase Reconstruction
0.1
−0.1
0 0.1 0.2 0.3 0.4

Time (s)
(b)
1
Input Amplitude Spectrum
Reconstructed Amplitude Spectrum
0.75
0.5
0.25
0
10 30 50 70 90 110
Frequency (Hz)
(c)
2 Input Phase Spectrum
Reconstructed Phase Spectrum
1
Phase (rad)
−1
−2
10 30 50 70 90 110
Frequency (Hz)
FIG. 1. Case A: Minimum phase input. (a) Input signal and minimum phase reconstruction gen-
erated using wavemin and minrceps respectively. (b) Amplitude spectra of input and output. (c)
Phase spectra of input and output. The curves are so close in value that they appear to lie on top
of one another.

(a) (b)
1 Input Wavelet 1 Reconstructed Wavelet
0.5 0.5
0 0
−0.5 −0.5
−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
FIG. 2. Case A: Pole-zero plots showing location of poles and zeroes of signal. Poles are de-
noted by x’s and zeros by o’s. (a) Input signal. (b) Minimum-phase reconstruction using minrceps
algorithm.
3
Absolute difference (10−3 )
Absolute Difference
2.5
2
1.5
1
0.5
0 0.1 0.2 0.3 0.4

Time (s)
FIG. 3. Case A: Absolute difference between output signal shown in Figure 1a and a minimum
phase reconstruction of the minrceps ouput using a Hilbert transform reconstruction method.
Mixed (linear) phase input (Case B)
The second test case has the same amplitude spectrum as Case A, but we apply a linear
phase to delay the energy of the signal - giving a mixed phase input. Figure 4a shows
that the algorithm has shifted the energy of the output signal to the beginning of the signal
with an identical shape to the minimum phase input in 1a. We also note no change in the
amplitude spectrum, which is desired property (Figure 4b). In Figure 4c, we observe the
linear phase of the input vs. the phase spectrum of the minimum phase reconstruction which
is near zero in comparison. The pole-zero plots in Figure 5 show that the minrceps output
has visibly placed the zeroes inside the unit cirlce, making the output minimum phase by
the pole-zero definition. Note that the pole-zero plot for the reconstructed signal in Case B
(Figure 5b) is slightly different from either pole-zero plot in Case A (Figure 2). Looking at
the difference between our minrceps output and a Hilbert transform reconstruction of itself
in Figure 6, we see that the absolute difference on average between the two is actually
smaller than in Case A by about an order of magnitude. This is an interesting result,
but acceptable nonethless as errors are still very small between the minrceps and Hilbert
transform reconstructions.

Smith & Ferguson
(a)
0.2
0.1
0
−0.1
Input Signal
−0.2 Minimum Phase Reconstruction
0 0.1 0.2 0.3 0.4
Time (s)
(b)
1
0.75
0.5
0.25
0
10 30 50 70 90 110
Frequency (Hz)
(c)
200
Input Phase Spectrum
150 Reconstructed Phase Spectrum
Phase (rad)
100
50
−50
10 30 50 70 90 110
Frequency (Hz)
FIG. 4. Case B: Mixed (linear) phase input. (a) Input signal generated using amplitude spectrum
from Figure 1b and applying a linear phase. (b) Amplitude spectra of input and output. (c) Phase
spectra of input and output.
(a) (b)
Input Wavelet 1 Reconstructed Wavelet

1
0.5 0.5
0 0
−0.5 −0.5
−1
−1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
FIG. 5. Case B: Pole-zero plots showing location of poles and zeroes of signal. Poles are denoted
by x’s and zeros by o’s. (a) Input signal. (b) Minimum phase reconstruction using minrceps.


Absolute Difference
0
0 0.1 0.2 0.3 0.4
Time (s)
FIG. 6. Case B: Absolute difference between output signal shown in Figure 4a and a minimum
phase reconstruction of the minrceps ouput using the Hilbert transform.
Mixed (linear) phase input with isolated zeroes (Case C)
The final two test cases aim to test the issue of zeroes in the amplitude spectrum of the
input. In Case C, we begin with the same amplitude and phase spectra as Case B (Figure
4), but add five isolated zeroes to the amplitude spectrum of the input (Figure 7). We see
in Figure 7a that the minrceps output has front-loaded energy, although there is now later-
time ringing. Looking at the amplitude spectra of the input and the output, we note that
the algorithm has begun to fill in the zeroes. The notches in the output at about 10, 50, 85,
and 95 Hz are all now non-zero, with no effect on the zero at 40 Hz. A new notch at about
30 Hz is created that was not in the input. The pole-zero plots in Figure 8 show that the
minrceps output appears to be minimum-phase by the classic definition, with all poles and
zeroes within the unit circle. Looking at the difference between our minrceps output and
a Hilbert transform reconstruction in Figure 9, we see larger absolute differences by about
an order of magnitude than in Cases A or B.

Smith & Ferguson
(a)
0.2
0.1
−0.1
Input Signal
−0.2 Minimum Phase Reconstruction
0 0.1 0.2 0.3 0.4
Time (s)
(b)
1
0.75
Zeroes
0.5
0.25
0
10 30 50 70 90 110
Frequency (Hz)
(c)
200
Phase (rad)
100
50
−50
10 30 50 70 90 110
Frequency (Hz)
FIG. 7. Case C: Mixed (linear) phase input with isolated zeroes. (a) Input signal generated using
phase spectrum from Figure 4c, and amplitude spectrum from Figure 1b with a number of isolated
zeros introduced. (b) Amplitude spectra of input and output, with isolated zeroes in input spectrum
shown in magenta circles. (c) Phase spectra of input and output.
(a) (b)
Input Wavelet 1 Reconstructed Wavelet

1
0.5 0.5
0 0
−0.5 −0.5
−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
FIG. 8. Case C: Pole-zero plots showing location of poles and zeroes of signals. Poles are denoted


6 Absolute Difference
0
0 0.1 0.2 0.3 0.4
Time (s)
FIG. 9. Case C: Absolute difference between output signal shown in Figure 7a and a minimum
phase reconstruction of the minrceps ouput using a Hilbert transform reconstruction.
Mixed (linear) phase input with interval of zeroes (Case D)
The final test, Case D, is an extension of Case C but with an interval of zeroes in
the input amplitude spectrum instead of isolated zeroes (Figure 10b). This case gives the
worst result visually, with the energy in the output appearing less front-loaded than any
of the other cases along with the ringing observed in Case C (Figure 10a). When we
look at the amplitude spectra, we see that the algorithm has again compensated for the
zeroes. There is still an interval from about 12-15 Hz in the output that is still near-zero.
This is likely the result for the visible worse result. Looking at the difference between
our minrceps reconstruction and a Hilbert transform reconstruction in Figure 12, we see a
similar magnitude of error as that with the isolated zeroes in Case C. Even with the visibly
worse result, we note that the pole-zero plots in Figure 11 show that all poles and zeroes
appear to still be within the unit circle.

Smith & Ferguson
(a)
0.2
0.1
−0.1
Input Signal
Minimum Phase Reconstruction
−0.2
0 0.1 0.2 0.3 0.4
Time (s)
(b)
1
0.75
Interval of Zeroes
0.5
0.25
0
10 30 50 70 90 110
Frequency (Hz)
(c)
200
Phase (rad)
100
50
−50
10 30 50 70 90 110
Frequency (Hz)
FIG. 10. Case D: Mixed (linear) phase input with interval of zeros. (a) Input signal generated using
phase spectrum from Figure 4c, and amplitude spectrum from Figure 1b with an interval of zeros
introduced. (b) Amplitude spectra of input and output, with interval of zeroes in input spectrum
shown in magenta circles. (c) Phase spectra of input and output.

(b)
1 Reconstructed Wavelet
(a)
0.5
1 Input Wavelet
0.5
0 0
−0.5
−1
−0.5
−1
−0.50 0.5 1
−1
−1 −0.5 0 0.5 1
FIG. 11. Case D: Pole-zero plots showing location of poles and zeroes of signals. Poles are denoted
Absolute Difference
6
0
0 0.1 0.2 0.3 0.4
Time (s)
FIG. 12. Case D: Absolute difference between output signal shown in Figure 10a and a minimum
phase reconstruction of the minrceps ouput using a Hilbert transform reconstruction.
Algorithm Performance
In cases A-D, we tested the performance of the minrceps algorithm which was used to
convert the signals to a minimum-phase equivalent with the same amplitude spectrum. We
observed in the case of a minimum-phase input (Case A), the algorithm essentially does
not change the signal. In Case B, we saw that the algorithm sucessefully converted a linear
phase input with a well-behaved amplitude spectrum to minimum phase. Even when we
added zeroes to the amplitude spectrum with linear phase in an isolated sense (Case C),
the result appears very similar to the output from Case B. When applying an interval of
zeroes to the amplitude spectrum with linear phase input in Case D, we get a result that
does not appear visibly to have as much of its energy front-loaded as the other three cases.
However, a pole-zero plot shows that it still appears to nearly satisfy the classic definition
of minimum phase. When we zero out parts of the input amplitude spectra as in cases C
and D, we observe ringing as well as the output amplitude spectrum not matching the input.
The ringing in the output is likely due to the zeroing of the input amplitude spectra and is
not caused by the application of the algorithm itself.
In addition to testing different cases where we have signals with ill-conditioned ampli-

Smith & Ferguson
tude spectra, we also need to consider the effect of altering or modifying the parameters α
and L. This could be especially important for improving results with problematic spectra,
such as Case D.
DECONVOLUTION TEST
Since we want to use the minrceps algorithm for analysis of geophysical and related
signals, we test its suitability in generating minimum phase outputs by a simple deconvo-
lution test. We take each minimum-phase reconstruction from the previous section as a
source wavelet (Figure 13a) and convolve it with a synthetically-generated white reflectiv-
ity (Figure 13b). We then perform a simple stationary Wiener deconvolution to return the
reflectivity where we use the input trace itself as the design trace, number of lags equal to
the length of the wavelet, and a stability factor of 0.1. In Figure 14, we show the resulting
seismic traces after convolving the wavelets in Figure 13a with the reflectivity in Figure
13b. The estimated reflectivities from deconvolution for each case along with the original
reflectivity are displayed in Figure 15.
(a)
0.2
Case A
Case B
0.1
Case C
Case D
0
−0.1
−0.2
0 0.1 0.2 0.3 0.4
Time (s)
(b)
1
Normalized Reflectivity
0.5
−0.5
−1
0.5 1 1.5 2 2.5 3
Time (s)
FIG. 13. Deconvolution test inputs. (a) Source wavelets which are outputs from Cases A-D from
the previous section, calculated using the minrceps function. (b) Synthetic reflectivity.
We see in Figure 15 that the returned reflectivities for all four cases are very similar.
Most significantly, there is not a large noticable difference in the recovered reflectivity from
Case A vs. Case D even though the wavelets appear considerably different (Figure 13a).
This suggests that the minrceps algorithm is returning an output that is close enough to
meeting the minimum-phase condition to give good deconvolution results, even though its

amplitude spectrum was ill-conditioned.
(a)
0.2 Case A
0.1
−0.1
−0.2
0.5 1 1.5 2 2.5 3
Time (s)
(b)
0.2 Case B
0.1
−0.1
−0.2
0.5 1 1.5 2 2.5 3
Time (s)
(c)
0.2
Case C
0.1
−0.1
−0.2
0.5 1 1.5 2 2.5 3
Time (s)
(d)
0.2
Case D
0.1
−0.1
−0.2
0.5 1 1.5 2 2.5 3
Time (s)
FIG. 14. Seismic traces generated convolving wavelets in Figure 13a with reflectivity in Figure 13b.

Smith & Ferguson
(a)
1
Normalized Reflectivity
0.5
−0.5
−1
0.5 1 1.5 2 2.5 3
Time (s)
(b)
Case A
Estimated reflectivity
0.2
−0.2
0.5 1 1.5 2 2.5 3

Time (s)
(c)
Case B
0.2
−0.2
0.5 1 1.5 2 2.5 3

Time (s)
(d)
Case C
0.2
−0.2
0.5 1 1.5 2 2.5 3

Time (s)
(e)
Case D
0.2
−0.2
0.5 1 1.5 2 2.5 3

Time (s)
FIG. 15. Original reflectivity (a) along with reflectivities estimated from stationary Wiener deconvo-
lution of the seismic traces in Figure 14 (b-e).

MINIMUM PHASE METRIC
We have shown the minrceps algorithm gives signals close enough to minimum phase to
get good deconvolution results. Since we are looking at signals that can only be described
minimum phase approximatley, we propose two measures to quanitatively describe their
"minimum phaseness", or how close they are in fact to being minimum phase. Eventually,
we hope to gain some idea of how accurate (or inaccurate) a certain geophysical process,
like deconvolution is based on how strongly our source wavelet violates the minimum-
phase criteron. Our methods originate with the systems definition of minimum phase,
namely poles and zeroes. Our metrics take into account both the number of poles and
zeroes on or outside the unit circle, and the average distance from the unit circle of those
poles and zeros that are located outside it.
Pole-Zero Ratio
We call our first metric the Pole-Zero Ratio (PZR). It is simply the ratio of the number
of zeroes and poles of a system/signal that lie on or outside the unit circle on the complex
plane to the total number of zeroes and poles. For example, a system with a poles at
z = ±2i and a zero at z = −0.5 would have a PZR of 2/3, because two-thirds of the total
poles and zeroes of the system lie on or outside the unit circle |z| = 1. It follows from this
that a minimum-phase signal/system would have a PZR of 0, becuase all poles and zeros by
definition are inside the unit circle. The corollary of this is that maximum-phase sequences
will have a PZR of 1, because all of their poles/zeros either lie on or outside the unit circle.
We would assume that the smaller the PZR, the closer the signal is to minimum phase.
Pole-Zero Distance
Our second metric is named the Pole-Zero Distance (PZD). It is a measure of the aver-
age distance of all of the poles and zeroes on or outside the unit circle from the unit circle
itself. For example, if all of the poles and zeroes of a sequence were located on the unit
circle, the PZD would return a value of zero. If all of the poles and zeroes are located inside
the unit circle (minimum phase), we give it a dummy value of −1. In the previous example
of two poles at z = ±2i and a zero at z = −0.5, we would calculate a PZD of 1,
P
(|P/Z|≥1) (|P/Z| − 1) (2 − 1) + (2 − 1)
P ZD = = = 1.
(Number of P/Z outside |z| = 1) (2)
With the PZD, we would potentially expect a larger number to represent a larger devi-
ation from minimum phase. However, this measure may be skewed by isolated poles and
zeroes lying well outside the unit circle.
Metric test
We apply these metrics to the four cases looked at previously: minimum-phase input,
linear phase input, linear phase with isolated zeroes in amplitude spectrum, and linear phase

Smith & Ferguson
with interval of zeroes (Figures 1a), 4a), 7a), and 10a) respectively. The results for the PZR
and PZD calculations for each case, as well as ratios for the outputs from the minrceps
algorithm are listed in Table 2.
Input PZR PZD Output PZR PZD

Case A 0 −1 0 −1
Case B 0.7559 0.0855 0 −1
Case C 0.7559 0.0493 0 −1
Case D 0.7402 0.0568 0.0472 0.0020
Table 2. Values of PZR and PZD calculated for the inputs and outputs used to test the minrceps
algorithm.
We see that in the Case A input, we have a PZR of zero and a PDZ of −1. This is
what we would expect, as Case A represents a minimum-phase wavelet. There is a drop
in the PZR in Case B, where about 25% of the poles/zeroes now lie on or outside the
unit circle. Similar values can be seen for Cases C and D. This drop is a result mainly of
applying a linear phase to the data. It is interesting to note that the addition of isolated
zeroes to the amplitude spectrum as in Case C did not lower the PZR from case B (with
a complete amplitude spectrum), but introducing an interval of zeroes as in Case D did
cause a slight drop in the PZR (making it closer to being minimum phase). With regards
to PZD, we notice that the Case B input has the highest value. If we compare Figures 5a
and 8a, we notice that although there are several poles in Case C that are further out from
the unit circle, the cluster of poles outside in Case B is on average further away. In our
output signals, we notice that Cases A-C are all exactly minimum phase and Case D is very
close, likely with one or two zeroes still on or outside the unit circle. From this limited
analysis, it is difficult to determine which metric better indicates "minimum phaseness".
More tests are required along with a more robust quantitative mesasurement of how "good"
our deconvolution results are to determine this.
CONCLUSIONS
In this report, we examined an algorithm to convert a signal of arbitrary phase to its

minimum phase equivalent. The minrceps algorithm is based upon calculation of the real
cepstrum, and has advantages to other Hilbert transform based minimum phase conversion
algorithms. We tested the performance of this algorithm on four test cases, incorporating
ill-conditioned inputs. The algorithm appears to be able to handle these. The outputs were
all found to very close to minimum phase by the classic definition from digital filtering. A
simple stationary deconvolution test reveals that the outputs were all close enough to mini-
mum phase that statistical deconvolution was able to return relatively accurate estimates of
reflectivity. Finally, we introduced two simple metrics to measure quanitiatively how close
a signal is to being minimum phase: the Pole-Zero Ratio (PZR) and the Pole-Zero Distance
(PZD), and preliminary calculation of these in the test cases indicates more analysis is re-
quired to get a better measure of how close a signal is to being minimum phase using these
metrics.

Future Work
The minrceps algorithm is an alternative to other Hilbert transform minimum-phase

reconstruction methods, and there is a possibility that the process can be linearized as it
does not require a separation and recombination of the phase and amplitude spectra of the
signal in question. We would like to investigate this possibility. Also, we need to develop
a better metric to quantitatively assess the deconvolution results vs. an input reflectivity.
Finally, in relation to the aforementioned desire to better quantify deconvolution results,
we would like to analyze deconvolution performance vs. PZR and PZD to better constrain
which is a better measure of minimum phaseness.
ACKNOWLEDGEMENTS
We thank the sponsors of CREWES for their support. We also gratefully acknowl-
edge support from NSERC (Natural Science and Engineering Research Council of Canada)
through the grant CRDPJ 379744-08.

Smith & Ferguson
REFERENCES
Benesty, J., Sondhi, M. M., and Huang, Y., 2008, Springer handbook of speech processing: Springer.
Bogert, B. P., Healy, M. J. R., and Tukey, J. W., 1963, The quefrency analaysis of time series for echoes:
Cepstrum, pseudo autovariance, cross-cepstrum and saphe cracking, in Rosenblatt, M., Ed., Time Series
Analysis, chap. 15, Wiley, 209–243.
Bourqui, J., Sill, J. M., and Fear, E. C., 2012, A prototype system for measuring microwave frequency
reflections from the breast: Journal of Biomedical Imaging, 2012, 9.
Eaton, J. W., Bateman, D., Hauberg, S., and Wehbring, R., 2013, GNU Ocatave: A high-level interactive
language for numerical computations - Edition 3 for Octave version 3.8.0.
Fear, E. C., Li, X., Hagness, S. C., and Stuchly, M. A., 2002, Confocal microwave imaging for breast cancer
detection: Localization of tumors in three dimensions: Biomedical Engineering, IEEE Transactions on,
49, No. 8, 812–822.
Gurbuz, A. C., McClellan, J. H., and Scott, W. R., 2009, A compressive sensing data acquisition and imaging
method for stepped frequency gprs: Signal Processing, IEEE Transactions on, 57, No. 7, 2640–2650.
Herrmann, O., and Schuessler, W., 1970, Design of nonrecursive digital filters with minimum phase: Elec-
tronics Letters, 6, No. 11, 329–330.
Jol, H. M., 2008, Ground penetrating radar theory and applications: Elsevier.
Karl, J. H., 1989, An Introduction to Digital Signal Processing: Academic Press Professional, Inc.
Lamoureux, M. P., Gibson, P. C., and Margrave, G. F., 2011, Minimum phase and attenuation models in
continuous time: CREWES Research Report, 23.
Lamoureux, M. P., and Margrave, G. F., 2007a, An analytic approach to minimum phase signal: CREWES
Research Report, 19.
Lamoureux, M. P., and Margrave, G. F., 2007b, A band-limited minimum phase calculation: CREWES
Lamoureux, M. P., and Margrave, G. F., 2007c, A minimum-phase, band-limited delta spike: CREWES
Margrave, G. F., 2001, Numerical methods of exploration seismology with algorithms in matlab: The Uni-
versity of Calgary Publ.
Margrave, G. F., 2013, Goph 517 - methods of seismic data processing: course lecture notes.
Margrave, G. F., Lamoureux, M. P., and Henley, D. C., 2011, Gabor deconvolution: Estimating reflectivity
by nonstationary deconvolution of seismic data: Geophysics, 76, No. 3, W15–W30.
Mian, G. A., and Nainer, A., 1982, A fast procedure to design equiripple minimum-phase fir filters: Circuits
and Systems, IEEE Transactions on, 29, No. 5, 327–331.
Oppenheim, A. V., 1965, Superposition in a class of nonlinear systems: Ph.D. thesis, Massachusetts Institute
of Technology.
Oppenheim, A. V., and Schafer, R. W., 2009, Discrete-Time Signal Processing: Prentice Hall Press.
Pei, S.-C., and Lin, H.-S., 2006, Minimum-phase fir filter design using real cepstrum: Circuits and Systems
II: Express Briefs, IEEE Transactions on, 53, No. 10, 1113–1117.
Quatieri, T. F., Jr, and Oppenheim, A. V., 1981, Iterative techniques for minimum phase signal reconstruction
from phase or magnitude: Acoustics, speech, and signal processing, IEEE Transactions on, ASSP-29,
No. 6, 1187–1193.

Robinson, E. A., 1967, Predictive decomposition of time series with application to seismic exploration: Geo-
physics, 32, No. 3, 418–484.
Robinson, E. A., and Treitel, S., 1967, Principles of digital wiener filtering: Geophysical Prospecting, 15,
No. 3, 311–332.
Sherwood, J., and Trorey, A., 1965, Minimum-phase and related properties of the response of a horizontally
stratified absorptive earth to plane acoustic waves: Geophysics, 30, No. 2, 191–197.
Stathaki, T., and Fotinopoulos, I., 2001, Equiripple minimum phase fir filter design from linear phase systems
using root moments: Circuits and Systems II: Analog and Digital Signal Processing, IEEE Transactions
on, 48, No. 6, 580–587.
Yilmaz, Ö., 2001, Seismic data analysis, vol. 1: Society of Exploration Geophysicists Tulsa.
Ziolkowski, A., and Bokhorst, K., 1993, Determination of the signature of a dynamite source using source
scaling, part 2: Experiment: Geophysics, 58, No. 8, 1183–1194.

Faadu

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Faadu

Uploaded by

Copyright:

Available Formats

Data analysis - processing and imaging

Minimum-phase signal calculation using the real cepstrum

The definition of a minimum-phase signal is physical in nature, defined as a signal

CREWES Research Report — Volume 26 (2014) 1

2 CREWES Research Report — Volume 26 (2014)

P C = |F −1 {ln |F{x(n)}|2 )}|2 , (1)

ĉ(n) = F −1 {Ĉ(eiw )} (4)

ĥ(n) = F −1 {Ĥ(eiw )} (5)

CREWES Research Report — Volume 26 (2014) 3

Minimum/maximum phase sequences and their complex cepstra

1. If x(n) is a minimum-phase sequence, its complex cepstrum ĉ(n) will be a causal

Relationship between minimum-phase sequence and its complex cepstrum

From this, if x(n) is a finite minimum-phase sequence, the summation in Equation 6 is

Reconstruction of a causal sequence from its even part

(Pei and Lin, 2006).

4 CREWES Research Report — Volume 26 (2014)

Fourier transform pair between time and frequency domain

xe (n) = F −1 {Re X(eiw ) }

xo (n) = F −1 {iIm X(eiw ) }

(Pei and Lin, 2006).

Relationship between complex cepstrum and real cepstrum

If x(n) is a real-valued sequence, its corresponding complex cepstrum is also real-

ĉe (n) = F −1 {Re[Ĉ(eiw )]} = F −1 {ln X(eiw ) }

ĉe (n) = F −1 {Ĥ(eiw )} = ĥ(n) (12)

IMPLEMENTATION OF MINRCEPS ALGORITHM

CREWES Research Report — Volume 26 (2014) 5

ĥ(n) = ĥmin (n). (14)

1. Choose a value α ≤ 1 and α ∼

4. Construct ĉα,min (n) from ĥα,min (n) using Equation 8.

6 CREWES Research Report — Volume 26 (2014)

Our four test cases are summarized as follows:

Case A: Minimum phase wavelet

Table 1. Parameters of input wavelets used to test minrceps algorithm

Minimum-phase input (Case A)

CREWES Research Report — Volume 26 (2014) 7

0 0.1 0.2 0.3 0.4

8 CREWES Research Report — Volume 26 (2014)

1 Input Wavelet 1 Reconstructed Wavelet

−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1

0 0.1 0.2 0.3 0.4

Mixed (linear) phase input (Case B)

CREWES Research Report — Volume 26 (2014) 9

Input Wavelet 1 Reconstructed Wavelet

10 CREWES Research Report — Volume 26 (2014)

Absolute difference (10−4 )

Mixed (linear) phase input with isolated zeroes (Case C)

CREWES Research Report — Volume 26 (2014) 11

Input Wavelet 1 Reconstructed Wavelet

12 CREWES Research Report — Volume 26 (2014)

Absolute difference (10−2 )

Mixed (linear) phase input with interval of zeroes (Case D)

CREWES Research Report — Volume 26 (2014) 13

14 CREWES Research Report — Volume 26 (2014)

CREWES Research Report — Volume 26 (2014) 15

16 CREWES Research Report — Volume 26 (2014)

amplitude spectrum was ill-conditioned.

CREWES Research Report — Volume 26 (2014) 17

0.5 1 1.5 2 2.5 3

0.5 1 1.5 2 2.5 3

0.5 1 1.5 2 2.5 3

0.5 1 1.5 2 2.5 3

18 CREWES Research Report — Volume 26 (2014)

MINIMUM PHASE METRIC

ĉe (n) = F −1 {Re[Ĉ(eiw )]} = F −1 {ln X(eiw )}